UC San Diego

In recent decades, there has been a tremendous amount of effort in the realm of brain-machine interfaces to develop neurally-driven prostheses able to restore lost function. A specific focus has centered on the engineering of communication prostheses aimed at restoring speech capabilities to people with vocal impairments. The study of human speech is challenging due to the absence of a suitable animal model capable of complex learned vocal behavior that facilitates the understanding of the brain circuitry involved in vocal production. Developing such a proxy model is crucial for understanding the neural basis of vocal behavior, for the informed design of neurotechnology aimed at restoring speech and for expediting clinical translation.

In this work, I propose a songbird model for vocal production. This thesis has four main contributions. Firstly, I introduce a novel surgical procedure that enables the simultaneous recording of neural populations in two regions of the zebra finch’s vocal-motor brain pathway using Neuropixels probes. Secondly, I use state-space analysis and latent-variable modeling techniques to highlight key differences in the neural dynamics that drive the population activity of the HVC and RA networks during song. These analyses provide new insights into their different functional roles during vocal production. Next, I investigate various approaches to continuously decode neural features into vocal output. The architectures proposed in this research are aimed at building vocal prostheses with unbounded outputs, bringing us closer to decoded vocal intent that resembles naturalistic behavior. Lastly, I explore new types of artificial neural networks that maintain performance in power-constrained environments.